A0706
Title: Latent position co-clustering for multiple social networks
Authors: Michael Fop - University College Dublin (Ireland) [presenting]
CJ Clarke - University College Dublin (Ireland)
Abstract: Social networks often involve multiple types of relationships, such as friendship, collaboration, and communication, resulting in a collection of networks recording different social dimensions over the same set of individuals. Analyzing such data requires methods that can uncover structure within single networks and across different relational dimensions. A latent position co-clustering model is introduced that jointly clusters networks and their constituent nodes. Built on a hierarchical mixture-of-mixtures formulation, the model simultaneously performs dimension reduction and two-level clustering. At the network level, it groups networks that share similar latent topological structures. At the node level, it uncovers local connectivity patterns such as communities or social roles. The latent space provides a parsimonious and interpretable representation of both global and local structure. The model adopts a Bayesian nonparametric framework based on mixtures of finite mixtures, which place priors on the number of mixture components at both levels and incorporate sparse priors to encourage parsimonious clustering. Inference is conducted via Markov chain Monte Carlo, employing a telescoping sampling strategy and a tailored post-processing procedure. Applications to real-world social multiplex data reveal interpretable network-level clusters aligned with contextual features, and node-level clusters that reflect roles and social patterns.
